1 一般优化问题
1.1 Standard Form 标准型
min
f
0
(
x
)
x
∈
R
n
s
.
t
.
f
i
(
x
)
≤
0
,
i
=
1
,
⋯
,
m
,
h
j
(
x
)
=
0
,
j
=
1
,
⋯
,
l
,
\begin{aligned} \min \quad & f_0(\boldsymbol x) \quad \boldsymbol x \in \mathbb R^n \\ \mathrm{s.t.} \quad & f_i(\boldsymbol x) \leq 0, \quad i=1,\cdots,m, \\ & h_j(\boldsymbol x) = 0, \quad j =1,\cdots, l, \end{aligned}
mins.t.f0(x)x∈Rnfi(x)≤0,i=1,⋯,m,hj(x)=0,j=1,⋯,l,
优化变量:
x
=
(
x
1
,
.
.
.
x
n
)
x=(x_1,...x_n)
x=(x1,...xn)
目标函数:
f
0
:
R
n
→
R
f_0: R^n\to R
f0:Rn→R
不等式约束:
f
i
:
R
n
→
R
f_i: R^n\to R
fi:Rn→R
等式约束:
h
i
:
R
n
→
R
h_i:R^n\to R
hi:Rn→R
最优值:
p
∗
=
i
n
f
{
f
0
(
x
)
∣
f
i
(
x
)
≤
0
,
i
=
1
,
.
.
.
m
,
h
i
(
x
)
=
0
,
i
=
1
,
.
.
.
,
l
}
p^*=inf\{f_0(x)|f_i(x)\leq 0,i=1,...m,h_i(x)=0,i=1,...,l\}
p∗=inf{f0(x)∣fi(x)≤0,i=1,...m,hi(x)=0,i=1,...,l}
最优解:
f
0
(
x
∗
)
=
p
∗
f_0(x^*)=p^*
f0(x∗)=p∗
隐式约束:
x
∈
D
=
⋂
d
o
m
f
i
∩
⋂
d
o
m
h
i
x\in D=\bigcap domf_i \cap \bigcap dom h_i
x∈D=⋂domfi∩⋂domhi
有时候我们只需要寻找一个可行点,
f
0
(
x
)
f_0(x)
f0(x)设置为常数。
1.2 stationary point 驻点
定义:
▽
f
(
x
)
=
0
\bigtriangledown f(x)=0
▽f(x)=0
局部极小:存在一个邻域内,使其在该邻域最小。
全局极小:显然
鞍点:存在一个邻域使得在改邻域内排在中间。
2 凸优化问题
2.1 标准形式
min
f
0
(
x
)
x
∈
R
n
s
.
t
.
f
i
(
x
)
≤
0
,
i
=
1
,
⋯
,
m
,
A
x
=
b
\begin{aligned} \min \quad & f_0(\boldsymbol x) \quad \boldsymbol x \in \mathbb R^n \\ \mathrm{s.t.} \quad & f_i(\boldsymbol x) \leq 0, \quad i=1,\cdots,m, \\ & Ax=b \end{aligned}
mins.t.f0(x)x∈Rnfi(x)≤0,i=1,⋯,m,Ax=b
目标函数是凸函数,定义在凸集上。
对于约束,凸函数的下水平集为凸集,凸集合的交为凸集。
为什么研究凸问题?凸问题的局部极小就是全局极小。
f
(
y
)
≥
f
(
x
∗
)
+
f
′
(
x
∗
)
T
(
y
−
x
∗
)
f(y)\geq f(x^*)+f'(x^*)^T(y-x^*)
f(y)≥f(x∗)+f′(x∗)T(y−x∗)
2.2最优性条件
对于可微函数
x
x
x是最优点 iff
f
′
(
x
∗
)
T
(
y
−
x
∗
)
≥
0
f'(x^*)^T(y-x^*)\geq 0
f′(x∗)T(y−x∗)≥0 for all feasible
y
y
y.
证明 见上面泰勒展开。
另外条件见后一章kkt.
2.3 等价
2.3.1 消除等式约束
min
f
0
(
F
z
+
x
0
)
x
∈
R
n
s
.
t
.
f
i
(
F
z
+
x
0
)
≤
0
,
i
=
1
,
⋯
,
m
,
\begin{aligned} \min \quad & f_0(Fz+x_0) \quad \boldsymbol x \in \mathbb R^n \\ \mathrm{s.t.} \quad & f_i(Fz+x_0) \leq 0, \quad i=1,\cdots,m, \\ \end{aligned}
mins.t.f0(Fz+x0)x∈Rnfi(Fz+x0)≤0,i=1,⋯,m,
where
F
F
F and
x
0
x_0
x0,
A
x
=
b
⟺
x
=
F
z
+
x
0
Ax=b \Longleftrightarrow x=Fz+x_0
Ax=b⟺x=Fz+x0 for some
z
z
z
x
0
x_0
x0是任意可行解,
R
(
F
)
=
N
(
A
)
R(F)=N(A)
R(F)=N(A)即
A
F
=
0
AF=0
AF=0
推广到变量替换。
2.3.2 对不等式引入松弛变量
min f 0 ( x ) x ∈ R n s . t . f i ( x ) + s i = 0 , i = 1 , ⋯ , m , s i ≥ 0 \begin{aligned} \min \quad & f_0(\boldsymbol x) \quad \boldsymbol x \in \mathbb R^n \\ \mathrm{s.t.} \quad & f_i(\boldsymbol x) +s_i =0, \quad i=1,\cdots,m, \\ & s_i\geq0 \end{aligned} mins.t.f0(x)x∈Rnfi(x)+si=0,i=1,⋯,m,si≥0
2.3.3 上镜图形式
min t x ∈ R n s . t . f 0 ( x ) ≤ t f i ( x ) ≤ 0 , i = 1 , ⋯ , m , A x = b \begin{aligned} \min \quad & t \quad \boldsymbol x \in \mathbb R^n \\ \mathrm{s.t.} \quad& f_0(\boldsymbol x) \leq t\\ \quad & f_i(\boldsymbol x) \leq 0, \quad i=1,\cdots,m, \\ & Ax=b \end{aligned} mins.t.tx∈Rnf0(x)≤tfi(x)≤0,i=1,⋯,m,Ax=b
2.3.4 降维 极小化其他变量
min i n f y f 0 ( x , y ) x ∈ R n s . t . f i ( x ) ≤ 0 , i = 1 , ⋯ , m , A x = b \begin{aligned} \min \quad & inf_y f_0(x,y) \quad \boldsymbol x \in \mathbb R^n \\ \mathrm{s.t.} \quad & f_i(\boldsymbol x) \leq 0, \quad i=1,\cdots,m, \\ & Ax=b \end{aligned} mins.t.infyf0(x,y)x∈Rnfi(x)≤0,i=1,⋯,m,Ax=b
3 拟凸问题
min
f
0
(
x
)
x
∈
R
n
s
.
t
.
f
i
(
x
)
≤
0
,
i
=
1
,
⋯
,
m
,
A
x
=
b
\begin{aligned} \min \quad & f_0(\boldsymbol x) \quad \boldsymbol x \in \mathbb R^n \\ \mathrm{s.t.} \quad & f_i(\boldsymbol x) \leq 0, \quad i=1,\cdots,m, \\ & Ax=b \end{aligned}
mins.t.f0(x)x∈Rnfi(x)≤0,i=1,⋯,m,Ax=b
f
0
f_0
f0是拟凸函数,
f
i
f_i
fi是凸函数。拟凸优化问题可以转为多个连续的凸问题。
4 凸优化问题分类
- LP,QP,SOCP,SDP
4.1 LP Linear Programming
min
c
T
x
+
d
x
∈
R
n
s
.
t
.
G
x
≤
h
A
x
=
b
\begin{aligned} \min \quad & c^Tx+d \quad \boldsymbol x \in \mathbb R^n \\ \mathrm{s.t.} \quad &Gx\leq h \\ & Ax=b \end{aligned}
mins.t.cTx+dx∈RnGx≤hAx=b
min
∣
∣
x
∣
∣
∞
||x||_{\infty}
∣∣x∣∣∞等价于
t
≤
x
≤
t
t\leq x \leq t
t≤x≤t ,min
t
t
t
min
∣
∣
x
∣
∣
1
||x||_{1}
∣∣x∣∣1等价于
−
t
i
≤
x
≤
t
i
-t_i\leq x \leq t_i
−ti≤x≤ti ,min
∑
t
i
\sum t_i
∑ti
4.2QP Quadratic Programming
min
1
2
x
T
P
x
+
q
T
+
r
x
∈
R
n
s
.
t
.
G
x
≤
h
A
x
=
b
\begin{aligned} \min \quad & \frac{1}{2}x^TPx+q^T+r \quad \boldsymbol x \in \mathbb R^n \\ \mathrm{s.t.} \quad &Gx\leq h \\ & Ax=b \end{aligned}
mins.t.21xTPx+qT+rx∈RnGx≤hAx=b
目标函数是二次,(
P
∈
S
+
n
P\in S_{+}^n
P∈S+n)约束为线性。
4.3 QCQP Quadratically Constrained QP
min
1
2
x
T
P
0
x
+
q
0
T
+
r
0
s
.
t
.
1
2
x
T
P
i
x
+
q
i
T
+
r
i
≤
0
A
x
=
b
\begin{aligned} \min \quad & \frac{1}{2}x^TP_0x+q_0^T+r_0 \\ \mathrm{s.t.} \quad &\frac{1}{2}x^TP_ix+q_i^T+r_i \leq 0 \\ & Ax=b \end{aligned}
mins.t.21xTP0x+q0T+r021xTPix+qiT+ri≤0Ax=b
P
i
∈
S
+
n
P_i\in S_{+}^n
Pi∈S+n
约束以及目标函数均为约束
4.4 SOCP Second-Order Cone Programming
min
f
T
x
s
.
t
.
∣
∣
A
i
x
+
b
i
∣
∣
≤
c
i
T
x
+
d
i
F
x
=
g
\begin{aligned} \min \quad & f^Tx \\ \mathrm{s.t.} \quad &||A_ix+b_i||\leq c_i^Tx +d_i \\ & Fx=g \end{aligned}
mins.t.fTx∣∣Aix+bi∣∣≤ciTx+diFx=g
目标函数为线性约束为二阶锥。范数锥
∣
∣
x
∣
∣
≤
t
||x||\leq t
∣∣x∣∣≤t
4.5 SDP Semidefinite Programming
min
c
T
x
s
.
t
.
∑
x
i
F
i
≼
G
A
x
=
b
\begin{aligned} \min \quad & c^Tx \\ \mathrm{s.t.} \quad &\sum x_iF_i\preccurlyeq G \\ & Ax=b \end{aligned}
mins.t.cTx∑xiFi≼GAx=b
线性矩阵不等式。